On 04/09/2018 07:20 AM, Oleg Nesterov wrote:
> On 04/04, Waiman Long wrote:
>> --- a/kernel/locking/percpu-rwsem.c
>> +++ b/kernel/locking/percpu-rwsem.c
>> @@ -179,8 +179,10 @@ void percpu_up_write(struct percpu_rw_semaphore *sem)
>>
>> /*
>> * Release the write lock, this will allow readers back in the game.
>> + * percpu_up_write() may be called from a task different from the one
>> + * taking the lock.
>> */
>> - up_write(&sem->rw_sem);
>> + up_write_non_owner(&sem->rw_sem);
>>
>> /*
>> * Once this completes (at least one RCU-sched grace period hence) the
>> diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
>> index 30465a2..140d5ef 100644
>> --- a/kernel/locking/rwsem.c
>> +++ b/kernel/locking/rwsem.c
>> @@ -222,4 +222,17 @@ void up_read_non_owner(struct rw_semaphore *sem)
>>
>> #endif
>>
>> +#ifdef CONFIG_DEBUG_RWSEMS
>> +/*
>> + * release a write lock from a different task
>> + */
>> +void up_write_non_owner(struct rw_semaphore *sem)
>> +{
>> + rwsem_release(&sem->dep_map, 1, _RET_IP_);
>> + DEBUG_RWSEMS_WARN_ON(!sem->owner || (sem->owner == RWSEM_READER_OWNED));
>>
>> + rwsem_clear_owner(sem);
>> + __up_write(sem);
>> +}
> Hmm. Can you look at lockdep_sb_freeze_release() and lockdep_sb_freeze_acquire()?
These 2 functions are there to deal with the lockdep code.
> At first glance, it would be much better to set sem->owner = current in
> percpu_rwsem_acquire(), no?
The primary purpose of the owner field is to enable optimistic spinning
to improve locking performance. So it needs to be set during an
up_write() call.
My rwsem debug patch does use it also to check for consistency in the
use of lock/unlock call. Anyway, I don't think it is right to set it
again in percpu_rwsem_acquire() if there is no guarantee that the task
that call percpu_rwsem_acquire will be the one that will do the unlock.
I am wondering if it makes sense to do optimistic spinning in the case
of percpu_rwsem where the unlocker may be a different task. We could set
a special code for writer owned lock, but don't do optimistic spinning
in this case.
Cheers,
Longman
